Development of a mid-infrared sensor system for early fire identification in cotton harvesting operations

Abstract

To realize early fire identification in cotton harvesting operations, a mid-infrared carbon monoxide (CO) sensor system was developed. To match the broadband light source with a 15° divergence angle, a multipass gas cell (MPGC) with an effective path length of 180 cm was designed to improve sensor sensitivity, leading to a limit of detection (LoD) of 0.83 parts-per-million by volume (ppmv). A damping module with springs at the bottom and front/back sides was fabricated, which can effectively reduce the vibration intensity by >80%. The sensor system can operate normally from −40 °C to 85 °C by stabilizing the temperature of the optical module through heating or cooling as well as using automotive electronic components. An adaptive early fire identification algorithm based on a dual-parameter threshold alarming method was proposed to avoid false and missing alarms. Field deployments on a harvester verified the good practicability of the sensor system.

---

1. Introduction

Cotton is an important cash crop and strategic material in China, with large yields and low production costs.^1–5 In order to improve cotton picking efficiency and provide comprehensive economic benefits, with the development of agricultural mechanization technology, cotton pickers have been widely used instead of manpower in cotton harvesting operations.^6–9 In the process of the cotton harvesting operation, the friction between the high-speed rotating cotton-picking head and the cotton stalk easily produces sparks, and these sparks are unavoidably wrapped in the cotton.^10,11 Consequently, early fire, namely smoldering, occurs inside the cotton, which is flameless and difficult to detect.¹² Once smoldering changes to flaming combustion and fire, great threats are posed to the safety and lives of people and property. Therefore, there is an urgent need to effectively realize early fire identification.

Common fire detection methods mainly include image recognition, and temperature and smoke alarms. As we know, gas as a product of fire appears earlier than a flame, temperature and smoke. So, gas sensors have potential application value in early fire detection.^13–15 Compared with traditional gas sensors, such as semiconductor sensors, electrochemical sensors, etc., gas sensors based on infrared absorption spectroscopy have attracted more and more attention because of their good selectivity and precision.^16,17 The gas products of cotton smoldering mainly include carbon dioxide (CO₂), carbon monoxide (CO) and some other harmful gases. Due to the presence of CO₂ in the environment, CO is selected as the characteristic gas for early fire identification in order to avoid interference from the background gas.^18–20 In addition, CO is also widely used in various trace gas detection scenarios, such as exhaled gas analysis,²¹ atmospheric monitoring,^22–25 combustion diagnosis,^26,27 and soil respiration analysis.²⁸ The detection sensitivity of parts-per-trillion by volume (pptv) levels can be obtained using a mid-infrared quantum cascade laser (MIR-QCL).^21,24,28 However, sensor systems based on the MIR-QCL are unsuitable for field deployment due to their large volumes and high cost, which make them unconducive to popularization and application.^23,25 Using a non-dispersive infrared absorption (NDIR) detection method, a CO sensor system was developed for early fire identification in cotton harvesting operations by considering both the performance and the cost. In recent years, some commercial NDIR sensors have been developed, such as the aSENSE MIII sensor from Seaseair, the Novagas2 sensor from Novasis, and the MicroSENS 180-HS sensor from Micro-Hybrid. A short optical path was specially designed for miniaturization, resulting in the low sensitivity of these sensors. In addition, these commercial sensors are expensive (up to thousands of dollars) and unsuitable for the popularization of agricultural applications. At the same time, some research studies have also been carried out, mainly focusing on improving the sensitivity, enlarging the optical path lengths, and reducing the power consumption and noise.^29–31 However, most of these sensors can only be operated in a laboratory or a good environment, and tend to perform poorly in harsh environments without any protection.

To realize early fire identification in cotton harvesting operations, a low-cost, high-sensitivity, anti-environmental interference sensor system was developed. A multipass gas cell (MPGC) with an effective path length of 180 cm was modeled and fabricated to increase the sensitivity of the sensor system to meet the need for early fire identification. Because of harsh harvesting operation sites, dust and moisture in the air entering the MPGC can affect the optical path structure. Since the sensor is placed on the harvester, large vibration and temperature differences severely influence the stability of the sensor system. Focusing on these problems in the harvesting operation, the developed sensor system was specially designed and optimized. A gas pretreatment module was used to filter out dust and moisture before the injection of ambient air into the MPGC to improve the measurement accuracy. A damping module with springs at the bottom and front/back sides of the sensor system was prepared to effectively reduce the vibration intensity and improve the stability. To cope with the influence of large temperature differences between day and night on the performance of the sensor system, a temperature control module was designed to stabilize the temperature of the optical module through heating or cooling and automotive electronic components were used.

Furthermore, most fire sensors only set one parameter (such as gas concentration, temperature, etc.) as the alarm threshold signal.³² However, once the monitored parameter is changed by external interference (such as vibration and temperature drifts), this detection method is then subject to a false alarm. In order to solve this problem, this work proposes an adaptive early fire identification algorithm based on the combination of the changing magnitude and the changing rate of CO concentration. The dual-parameter threshold alarming method can effectively identify early fire and avoid the occurrence of false alarms and missing alarms.

2. MIR CO sensor using a MPGC

2.1 System structure

As shown in Fig. 1, the whole system structure consists of an optical part and an electrical part. In the optical part, the emitted light from a light source (Hawkeye Technologies, model IR55) with an approximately 15° divergence angle was converged using a convex lens and then was reflected by a beam splitter into a MPGC. Subsequently, the incident light was reflected six times by two concave mirrors to form an effective optical path length of 180 cm. The concave mirrors had a focal length of 300 mm and a reflectivity of 98%. Then, the light was transmitted to a pyroelectric detector (InfraTec, model LMM-242-IH) through the beam splitter. Since the infrared light source and pyroelectric detector were susceptible to temperature, a temperature control module with an electric heating film (EHF) and a thermoelectric cooler (TEC) was used to adjust the internal temperature of the optical module. In the electrical part, a digital signal processor (DSP, Texas Instruments, model TMS320F28335) was used as the main control unit of the whole sensor system. A 4 Hz square wave signal generated from DSP was used to modulate the infrared light source with a light source driver. At the same time, the modulated optical signal was received and converted to two electrical signals by the pyroelectric detector. The two electric signals were amplified using a preamplifier to match the input range of the custom-made lock-in amplifier. The amplitude of the electric signals was extracted using the lock-in amplifier and then transmitted to the DSP. The DSP converted the amplitude to CO concentration by means of the calibrated and fitted relation. The information was transmitted between the DSP and the LabVIEW-based monitoring platform through a wireless network. Long-term and real-time monitoring of CO concentration was carried out using the monitoring platform and an early fire could be identified through a self-adaptive identification algorithm accompanied by a fire alarm.

Fig. 1 Structural diagram of the sensor system.

2.2 Detection theory

Different gases have specific absorption of light at different wavelengths, which conforms to the Lambert–Beer law,³³i.e.

I(λ) = I₀(λ) exp[−α(λ)cl] (1)

where I₀(λ) is the incident light intensity, I(λ) is the outgoing light intensity, α(λ) is the absorption coefficient, c is the gas concentration, and l is the optical path length. Let

I₀(λ) = φ(λ)i (2)

where i and φ(λ) are the driving current and the electro-to-optical conversion coefficient of the light source, respectively. During gas detection, the target gas is commonly mixed with dust and moisture, and then eqn (1) can be revised to

I(λ,l) = I₀(λ) exp[−α(λ)cl + m(λ)l + n(λ)l + p(λ)l] (3)

where m(λ) is the Raleigh scattering coefficient, n(λ) is the Mie scattering coefficient, and p(λ) is the absorption coefficient of H₂O.

Considering the measurement accuracy and system complexity, the dual-channel differential detection method based on the NDIR principle was adopted. The emitted light is absorbed by the target gas in the MPGC and then exits to the two sensing windows of the detector. The central wavelengths of the absorption channel and the reference channel are λ₁ and λ₂, respectively. The output signals U(λ₁,l) and U(λ₂,l) of the two corresponding channels can be expressed by eqn (4) and (5), respectively, as

where η(λ₁) and η(λ₂) are the optical-to-electro conversion coefficients at λ₁ and λ₂, respectively. k(λ₁) = exp[m(λ₁)l + n(λ₁)l + p(λ₁)l] and k(λ₂) = exp[m(λ₂)l + n(λ₂)l + p(λ₂)l]. The ratio between U(λ₁,l) and U(λ₂,l), (U(λ₁,l)/U(λ₂,l)), can be written as

Then, gas concentration c can be obtained by eqn (6), i.e.

2.3. Optical design

2.3.1. Light source and detector. The light source is an important part of the sensor system and some of its characteristics (including the stability of the optical power and modulation characteristics) can affect the limit of detection (LoD) and sensitivity of the sensor system. The used broadband light source is made with a thin thermosensitive conducting film of amorphous carbon, and infrared radiation is generated by heating this film.³⁴ The total optical power of the light source measured using an optical power meter was 4.38 mW and the optical power of the absorption channel was estimated to be 0.07 mW within the spectral range used for gas detection. As shown in Fig. 2(a), the light source was built into a Hawkeye parabolic optic to collimate the light, and the effective radiation area was 2.25 mm² with an approximately 15° divergence angle. A Fourier transform infrared spectrometer was used to measure the emission spectrum of the light source at different driving currents in air. The results show that the spectral range covers 2–20 μm, as shown in Fig. 2(b). In addition, the emission intensity at the same wavelength was different at different driving currents. During detection, the influence of the driving current change can be removed based on eqn (7). However, the ratio between φ(λ₁) and φ(λ₂), i.e., φ(λ₁)/φ(λ₂), is inconstant as the driving current changes, which still affects the detection result. So, the driving current should be constant. Therefore, a voltage-control constant current source circuit was developed based on the above considerations. When the driving current was set to 130 mA, the measurement result showed an averaged output current of 130.02 mA with a standard deviation of 0.01 mA.

Fig. 2 (a) The photo spectrum and (b) the emission spectrum of the light source.

The used detector is a dual-channel detector with an integrated operational amplifier circuit, a temperature compensation circuit and a filter circuit. The window for each channel has an optical filter. The filter for the absorption channel was centered at 4.66 μm with a half-power bandwidth (HPBW) of 180 nm. The filter for the reference channel was centered at 3.95 μm with a HPBW of 90 nm.³⁵ Considering the presence of H₂O and CO₂ in the air, Fig. 3 shows the simulated absorbance of 500 ppmv CO₂, 100 ppmv CO and 2% H₂O near 3.95 μm and 4.66 μm at a 300 K temperature, 1 atm pressure and 180 cm optical path based on the high-resolution transmission (HITRAN 2020) molecular absorption database.³⁶Fig. 3(a) shows that the absorbance of 100 ppmv CO is about five orders of magnitude larger than those of 500 ppmv CO₂ and 100 ppmv HCN within the spectral range of the absorption channel, so the interference of CO₂ and HCN can be negligible. The absorbance of 2% H₂O and 100 ppmv CO is on the same order of magnitude, so moisture should be removed before gas detection. In addition, 100 ppmv NO₂ and 100 ppmv NO had almost no absorption in the absorption channel. The absorbance of 500 ppmv CO₂, 2% H₂O, 100 ppmv CO, 100 ppmv HCN, 100 ppmv NO₂ and 100 ppmv NO within the spectral range of the reference channel was much smaller than the absorbance of 100 ppmv CO within the spectral range of the absorption channel. Therefore, the signal generated by the reference channel only represents the background noise.

Fig. 3 The absorbances of 500 ppmv CO₂, 2% H₂O, 100 ppmv CO, 100 ppmv HCN, 100 ppmv NO₂ and 100 ppmv NO near (a) 4.66 μm and (b) 3.95 μm.

2.3.2. Optical structure and MPGC design. In the harvesting environment, impurities such as dust can easily affect the performance of the optical path. Although the response time can be shortened using an open path detection scheme, it is at the expense of measurement accuracy. So, it is unsuitable to adopt an open optical path structure. Moreover, a single optical path structure design will considerably expand the volume of the sensor, which is also unsuitable for vehicle deployment. After comprehensive consideration, the MPGC is the best solution for the sensor. The propagation path of paraxial rays can be expressed by optical ray matrices. Any paraxial ray in a given cross-section can be represented by two coordinate parameters, namely the distance and the included angle between the ray and the optical axis. When the ray is above the optical axis, the included angle is positive, otherwise, it is negative. As shown in Fig. 4(a), two identical concave mirrors are placed in parallel at a distance of l. P₀ is the incident point on mirror M₁ and the distance from P₀ to the optical axis is d₀. The incident light S₀ reaches P₁ on mirror M₂ and reflected light S₁ is generated. The normal line is OP₁, the radius of curvature is r, and the incidence and reflection angles are both α. The angle between the incident light S₀ and the optical axis is θ₀ and the angle between the reflected light S₁ and the optical axis is θ₁. The distance between the P₁ and the optical axis is d₁. When r is relatively large, the light transmission distance can be approximately l. Both θ₀ and β are very small under the condition of paraxial rays, thus they can be considered as

Fig. 4 (a) The establishment of the light transmission matrix. (b) The light transmission path without the divergence angle. (c) The light transmission path with a 15° divergence angle. (d) The CAD image and photograph of the compact MPGC.

After the incident light S₀ is reflected, the propagation direction of the reflected light S₁ is changed to be below the optical axis, so θ₁ is a negative value, i.e.

θ₁ = −(α + β) = −(θ₀ + 2α) (10)

The position of the reflection point P₁ on M₂ can be described by eqn (11), i.e.

when r = 2l, eqn (11) can be simplified to

After six reflections, the position of the reflection point can be described by eqn (13), i.e.

It can be seen from eqn (13) that the position of the reflection point is the same as that of the incident point P₀.

In order to verify the above-mentioned optical ray matrices, a MATLAB program was designed for light propagation simulation. Fig. 4(b) shows the optical propagation path of the incident light without the divergence angle. Fig. 4(c) shows the optical propagation path of the incident light with a 15° divergence angle. In addition, it can be found that the light can return to the original incident position after 6 reflections. Fig. 4(d) shows the three-dimensional model and photograph of the prepared MPGC. The bottom and two sides of the MPGC are made of aluminum alloys, and the 30 cm-long cavity is made of acrylic. The radius of the 30 cm-long cavity was 4 cm, resulting in a volume of ∼1.5 L. Two gold-plated concave mirrors were embedded on both sides of the MPGC, one of which has a light entrance hole with a diameter of 1 cm. To ensure the airtightness of the MPGC, a flat window was used to seal the light entrance hole, and then the flat window was squeezed through an external sealing cover. Considering the miniaturization and compactness of the optical structure, the light source and detector were placed on the same side of the MPGC, and the light beam entered and exited the MPGC from the same light entrance hole. A gas inlet and a gas outlet are located on both sides of the MPGC, respectively.

2.4. Gas pretreatment module

Gas pretreatment can lessen the effect of dust and moisture on the sensor system and improve detection accuracy. The structural diagram of the gas pretreatment module is shown in Fig. 5(a), which mainly includes two coarse filters (CTM, model CFD-10-02, Japan), two fine filters (Swagelok, model SS-4FW-2, USA), a gas dryer (Perma Pure, model MD-110-48F-R, USA), a ball valve and a vacuum pump (KNF, model NMP830HP, Germany). A two-stage filtration composed of a coarse filter and a fine filter can effectively filter out impurities in the sample gas. Most of the large particulates with a diameter of >10 μm can be filtered out using the coarse filter, and the tiny particulates with a diameter of >2 μm can be filtered out using the fine filter. The combination of the ball valve and vacuum pump controls the gas flow rate together. Fig. 5(b) shows a photograph of the assembled gas pretreatment module.

Fig. 5 (a) Structural diagram and (b) photograph of the gas pretreatment module.

The gas dryer is a moisture exchanger composed of the Nafion polymer. The Nafion polymer can transfer water molecules from one side of the membrane to the other according to the difference in the humidity level on both sides of the membrane. With the assistance of the Nafion polymer, the gas dryer selectively removes water from the pumped gas sample. This selectivity for water allows the gas dryer to remove more moisture than other gas drying solutions while keeping most analytes in the gas sample unaffected. The sample gas and the purge gas flow inside and outside of the Nafion tubing, respectively, in opposite directions. The dry purge gas can be either external purge gas or reflux sample gas. The external purge gas sources mainly include nitrogen, instrument air and compressed air. However, this method might not be cost-effective for continuous or high-flow-rate applications. In this sensor, the dust and moisture are filtered out by the two-stage filtration and gas dryer in turn. Then, the sample gas ejected from the MPGC flows back into the purge gas flow path. Compared with the method using external purge gas, this method is simpler and cost-effective.

2.5. CO sensing performance evaluation

2.5.1. Sensor calibration and response time. A commercial gas mixing system was used to prepare 0 ppmv, 40 ppmv, 80 ppmv, 120 ppmv, 160 ppmv and 200 ppmv CO gas samples with a 1000 ppmv standard CO sample and high-purity nitrogen (N₂) using a dynamic gas distribution method. The prepared CO samples were injected into the MPGC in turn, and each gas sample was measured for 10 min. The value of U(λ₁,l)/U(λ₂,l) for each gas sample was averaged and plotted as a function of CO concentration. As shown in Fig. 6(a), the fitting curve indicates a relationship as follows:where the goodness of fit is 0.997.

Fig. 6 (a) The fitting curve of the voltage ratio U(λ₁,l)/U(λ₂,l) versus CO concentration. (b) The measurement result of the response time by injecting 0 ppmv, 100 ppmv and 0 ppmv CO samples into the MPGC, respectively.

The response time of the sensor is affected by several factors, such as the volume of the gas cell and the gas flow rate.³⁷ 0 ppmv and 100 ppmv CO gas samples were injected alternately into the MPGC at a gas flow rate of 1 L min⁻¹. The 10–90% response time was found to be ∼25 s for both rise and fall conditions, as shown in Fig. 6(b).

2.5.2. System stability. High-purity N₂ (i.e., 0 ppmv CO sample) was slowly injected into the MPGC under a standard ambient temperature and pressure for the stability test of the sensor system. Using N₂ as the target gas instead of the prepared gas sample can avoid the dilution error or noise introduced by the gas dilution process. The CO concentration was measured and recorded continuously for ∼22 min to evaluate the stability. As shown in Fig. 7(a), the 0 ppmv CO sample was found to be 0.03 ± 3.16 ppmv (1σ). The LoD of the sensor system was analyzed using Allan deviation. As shown in Fig. 7(b), an LoD of 0.83 ppmv for an averaging time of 0.25 s was achieved and the LoD was estimated to be 0.63 ppmv for an optimum averaging time of 58 s. The white Gaussian noise in the system can affect the limit of detection. Normally, the white Gaussian noise level decreases with increasing averaging time, so the Allan deviation decreases gradually. When the averaging time reaches a certain extent, Allan deviation is mainly affected by the system drift. At this point, the Allan deviation begins to increase. However, in this sensor, since filtering was used in data sampling and processing, the effect of increasing averaging time was weakened. Instead of decreasing it, system drift leads to a rise in the Allan deviation in the initial stage, as shown in Fig. 7(b).

Fig. 7 (a) The measured concentration of the 0 ppmv CO gas sample. (b) The variation curve of Allan deviation versus average time.

3. Enhancement of sensor stability

3.1. Stability enhancement under vibration

As shown in Fig. 8(a), the structure of the damping module is mainly composed of a base and four columns. The base is made from 304 stainless steel and supports the overall weight of the sensor system. The base is flanked by four columns, which are fixed on the instrument housing. There are four round holes dug in the base for placing springs. As shown in Fig. 8(b), when the instrument is subjected to external downward vibration, the springs at the bottom produce a reaction force to reduce the downward vibration. In addition, four springs were placed between the four columns and the recess to prevent the instrument from vibrating upward. Fig. 8(c) shows the outside view of the sensor system with the damping module. The shell of the whole sensor system adopts the IP67 degree of protection. The external 12 V power was supplied to the internal power module through an aviation plug.

Fig. 8 (a) Overall appearance structure diagram of the damping module. (b) Side view of the damping module with spring. (c) Outside view of the sensor system with the damping module. (d) Simulation results of vibration of the damping module based on MATLAB.

The sensor system with the damping module can be simplified to a single-degree-of-freedom dynamic model. The dynamic equation of the model satisfies eqn (15):

mx′′(t) + cx′(t) + kx(t) = f(t) (15)

where x(t) is the vibration displacement, f(t) is the external excitation, m is the weight of the sensor system, and c and k are the damping coefficient and stiffness of the spring, respectively. Further organizing eqn (15) yields eqn (16), i.e.

x′′(t) + 2wξx′(t) + w²x(t) = F(t) (16)

where w² = k/m, ξ = c/2mw, F(t) = f(t)/m. Duhamel's integral is a method to solve the response of the linear system in vibration theory. By solving eqn (16) with the Duhamel integral, the expression of vibration displacement x(t) can be obtained, i.e.where a and v₀ are the displacement and rate at t = 0, respectively, w_n is the natural frequency, and . The spring stiffness k is defined as the ratio of the load increment df to the deformation increment dλ, i.e.

The weight of the system m = 7 kg and the spring stiffness k = 700 kg m⁻¹. According to eqn (18), the initial displacement a = 0.01 m. Assume that the external excitation f(t) = 0.05 sin(t). Without considering the spring damping, the performance of the damping module was simulated based on MATLAB. The simulation results show that the vibration intensity can be reduced by ∼80% after using the damping module, which theoretically illustrates the practicality of the damping design, as shown in Fig. 8(d).

The developed sensor system equipped with a vibration sensor was used to carry out outdoor experiments to test the performance of the damping module. Fig. 9 shows the measured vibration intensity and the CO concentration in the atmosphere on an asphalt road, a grass road and a stone road under three different conditions. As shown in Fig. 9(a–c), there is a big difference in the vibration intensity in three road conditions. The average vibration intensities with the damping module under the three road conditions are reduced by 86%, 87% and 89%, respectively, compared with those of the sensor system without the damping module. Table 1 shows the detailed analysis results. The standard deviations of vibration intensity with the damping module are much lower than those without the damping module under the three road conditions. Therefore, the damping module can effectively absorb external vibration.

Fig. 9 The vibration intensity under the three road conditions: (a) asphalt road, (b) grass road and (c) stone road. The measured CO concentration in the atmosphere under the three road conditions: (d) asphalt road, (e) grass road and (f) stone road.

Table 1 Detailed analysis results of vibration intensity and CO concentration

Road conditions	Damping module	The average value of vibration intensity (mm s^−1)	Percentage reduction (%)	The fluctuation value of CO concentration (ppmv)	Percentage reduction (%)
Asphalt road	Without	4981	86	44.83	59
	With	689		18.41
Grass road	Without	7929	87	56.95	64
	With	998		20.50
Stone road	Without	6626	89	60.51	73
	With	717		16.59

---

In addition, to verify the beneficial impact of adding the damping module to the sensor system, the CO concentration in the atmosphere under the three road conditions mentioned above was measured using the developed sensor system. As shown in Table 1, the fluctuation of the CO concentration measured using the sensor system with the damping module on an asphalt road, a grass road and a stone road was reduced by 59%, 64% and 73%, respectively, compared with those without the damping module. Therefore, the damping module can effectively enhance both the stability and accuracy of the sensor system.

3.2. Stability enhancement under a large temperature change

To cope with the large temperature difference between day and night, both the electrical and the optical parts should operate normally within a large temperature variation range. In the electrical part, the electronic components used are all automotive grade, which ensures the normal operation of the electrical part under extreme temperature conditions. In the optical part, the pyroelectric detector used in the sensor system is sensitive to temperature, thus a temperature control module was designed to control the temperature of the optical module. Fig. 10(a) shows a photograph of the optical module, in which a PT100 thermal resistor was adhered to the beam splitter for measuring the internal temperature change of the optical module. As shown in Fig. 10(b), the housing of the temperature control module was designed by 3D printing technology to cover the optical module. As shown in Fig. 10(c), two TECs embedded at the top and side of the housing were used for cooling, and an EHF with a maximum power of 48 W installed inside the housing was used for heating. By comparing the real-time temperature with the target temperature, the DSP adopted a proportional-integral-derivative (PID) algorithm to keep the temperature constant by controlling the EHF to heat or the TEC to cool so as to minimize the impact of temperature change on the detector. Fig. 10(d) shows an internal photograph of the sensor system. The optical part and electrical part are arranged separately to achieve a small size and compact structure. The developed sensor system was put into a constant temperature and humidity test chamber (CTHTC, Komeg, model KM-225, China) to observe the performance of the temperature control module, as shown in Fig. 10(e).

Fig. 10 (a) Photograph of the optical module. (b) Three-dimensional model of the temperature control module. (c) Photograph of the temperature control module covering the optical module. (d) Internal photograph of the sensor system. (e) Photograph of the constant temperature and the humidity test chamber.

As shown in Fig. 11, a 0 ppmv CO sample (pure N₂) was injected into the MPGC, and the internal target temperature of the optical module was set to 21 °C. At the same time, the temperature of the CTHTC was adjusted from −40 °C to 85 °C and was maintained at −40 °C and 85 °C for 30 minutes, respectively. It took ∼35 minutes to rise from −40 °C to 85 °C. The external temperature of the optical module was always lower than the internal temperature in the first 46 minutes, which is mainly due to the continuous heating of the EHF. As the external temperature continued to rise and exceeded the internal target temperature, the TECs began to cool the internal environment of the optical module. The internal temperature kept rising slowly, finally reaching 34.06 °C. This was caused by the insufficient cooling efficiency of the TECs. The maximum cooling power of the TECs was 27 W. However, the measured CO concentration fluctuated between −9.71 ppmv and 9.95 ppmv during the whole process of external temperature change, benefiting from the temperature compensation of the pyroelectric detector within a certain temperature range. In future work, adding the water-cooling module to TECs can further increase cooling efficiency, leading to a further improvement in the stability of the sensor system.

Fig. 11 Performance test results of the temperature control module. The blue line represents the measured CO concentration, the red line represents the external temperature, and the black line represents the internal temperature of the optical module.

4. Self-adaptive early fire identification algorithm

The CO content in the atmosphere is very small. When the CO concentration increases rapidly in a short time, it can be considered that there may be a hidden fire. According to the changing magnitude and changing rate of CO concentration, a self-adaptive early fire identification algorithm was proposed, which can automatically adapt to the change of gas concentration in the background environment. Fig. 12 shows the execution flow chart of the algorithm. First, at the initial state, let i = 1 and j = 0. Then, record the CO background concentration in the environment for 30 s with a sampling rate of 4 Hz, leading to 120 samples. Calculate the average value a of the latest 120 samples to update the background concentration, i.e., a = (C_i−1 + C_i−2 + ⋯ + C_i−120)/120. Then, judge whether the difference b between C_i and a is larger than x, where x is the threshold of the changing magnitude. If yes, proceed to the next step; if not, let i = i + 1. Let the changing rate c = (C_i − C_i−120)/30, where C_i and C_i−120 are concentration values at 120 samples apart. Then, judge whether c is larger than the threshold y of the changing rate. If yes, let j = j + 1; if not, let j = 0 and i = i + 1. In order to eliminate the instantaneous rise of C_i and c caused by external factors, such as vibration, judge whether j is greater than 20, that is, the changing magnitude and rate of CO concentration exceed the set threshold values x and y within 5 consecutive seconds. If yes, start the alarm; if not, let i = i + 1, and continue to read the next concentration value.

Fig. 12 Execution flow chart of the self-adaptive early fire identification algorithm.

In order to determine the threshold x and y, 1 g of cotton was used for testing under smoldering conditions. An electric heating table was placed in a transparent acrylic box to heat the cotton. A special temperature setting ensured that the cotton did not produce an open flame and remained smoldering. The generated CO was extracted from the top of the transparent acrylic box and was pumped into the MPGC at a flow rate of 1 L min⁻¹. In addition, the gas pretreatment module was installed in front of the gas inlet of the MPGC to filter out dust and moisture. In the case of smoldering, the changing magnitude of CO concentration generated by 1 g of cotton over time is shown in Fig. 13(a). It can be found that the peak value is between 150 ppmv and 300 ppmv. Therefore, the threshold x of changing magnitude can be set to 100 ppmv. Fig. 13(b) shows the changing rate of CO concentration generated by 1 g of cotton over time. It can be found that the peak value exceeded 2 ppmv s⁻¹, so the threshold value y of the changing rate can be set to 2 ppmv s⁻¹. For broad applications of the sensor system, equal weights of plywood, wood and paper were used to carry out the smoldering experiment. The measurement results are similar to those of cotton smoldering, as shown in Fig. 13(a) and (b). Therefore, the same thresholds are also applicable to early fire identification for plywood, wood and paper. During detection, once the changing magnitude and the changing rate of CO concentration are higher than the set threshold x and y at the same time, the monitoring platform can distinguish them and then give an alarm.

Fig. 13 (a) The changing magnitude and (b) the changing rate of CO concentration generated by plywood, cotton, wood and paper with equal weights.

5. On-site deployment

The developed sensor system was deployed on a harvester to verify the impact in practical application. The sensor system, the gas pretreatment module and the battery were placed on one side of the cab, as shown in Fig. 14(a). In order to save space, the sensor system and the gas pretreatment module were stacked. The battery supplied power to the sensor system and the vacuum pump embedded in the gas pretreatment module. As shown in Fig. 14(b), one end of the tube was connected to the air inlet of the MPGC and the other end was put into the storage bin to extract the gas inside. The smoldering cotton was placed in the storage bin and the generated CO was pumped into the gas pretreatment module using the pump through the tube. The target gas entered the sensor system after being filtered and dried using the gas pretreatment module.

Fig. 14 (a) An internal photograph and (b) an external photograph of the sensor system installed on a harvester. (c) The magnitude and changing rate of CO concentration during the movement of the harvester.

The self-adaptive early fire identification algorithm was used to monitor the early fire in real time. Fig. 14(c) shows the magnitude and changing rate of CO concentration during the movement of the harvester, in which the pink area and the green area represent the judgment time and the alarm time, respectively. The background CO concentration in the air was recorded for a period of 30 s and the average value was determined to be 1 ppmv. At point A (t = 312 s), the cotton in the storage bin was ignited for smoldering. At point B (t = 342 s), the changing magnitude and changing rate of CO concentration were 100 ppmv and 2.98 ppmv s⁻¹, respectively, which exceeded the thresholds. At this time, the sensor system was continuously working and both the changing magnitude and changing rate of CO concentration exceeded the thresholds in the following 5 seconds. At point C (t = 347 s), the self-adaptive early fire identification algorithm identified the occurrence of an early fire and triggered sound and light alarms. At point D (t = 380 s), the changing magnitude and rate of CO concentration were 195 ppmv and 1.99 ppmv s⁻¹, respectively. Since the changing rate was lower than the set threshold of 2 ppmv s⁻¹, the alarm was terminated. The alarm time lasted for 33 s, which verified the practicability of the self-adaptive early fire identification algorithm. It should be noted that the CO produced by the harvester is different from that produced by crop combustion in terms of the changing magnitude and changing rate, which can be distinguished by the self-adaptive early fire identification algorithm. So, the harvester caused no effect on early fire identification throughout the experiment.

6. Conclusions

Focusing on the early fire hazard in cotton harvesting operations, a mid-infrared sensor system was developed for early fire identification. A gas pretreatment module was developed to decrease the impact of dust and moisture. A damping module was attached to the sensor system and outdoor experiments were implemented to verify its performance. The fluctuation of the measured CO concentration on an asphalt road, a grass road and a stone road was reduced by 59%, 64% and 73%, respectively, using the damping module. To cope with the influence of a large temperature difference between day and night on the performance of the sensor, a temperature control module was designed to control the temperature of the optical module. Pure N₂ was injected into the MPGC placed in the CTHTC. The measured CO concentration fluctuated between −9.71 ppmv and 9.95 ppmv during a change in the external temperature from −40 °C to 85 °C. However, the internal temperature of the optical module increased from 21 °C to 34.06 °C, which was caused by the insufficient cooling efficiency of the TECs. In future work, adding a water-cooling module to TECs can further increase the cooling efficiency. In view of the false alarm and missing alarm of the single-parameter warning mechanism, an adaptive early fire identification algorithm based on the combination of the changing magnitude and changing rate of the CO concentration was proposed. Through the smoldering experiment of cotton, the threshold of the changing magnitude and the threshold of the changing rate were determined as 100 ppmv and 2 ppmv s⁻¹, respectively. Related smoldering experiments confirmed that the thresholds are also applicable to the early fire identification of plywood, wood and paper. The field deployment and experiment on the harvester verified the good application value of the developed sensor system in early fire identification.

Author contributions

Yafei Li: conceptualization, methodology, data curation, and writing – original draft. Yang Lu: investigation and validation. Chuantao Zheng: conceptualization, investigation, writing – review and editing, and funding acquisition. Shuo Yang: conceptualization and validation. Kaiyuan Zheng: investigation and validation. Fang Song: investigation and data curation. Chunguang Li: writing – review and editing. Weilin Ye: conceptualization and methodology. Yu Zhang: writing – review and editing. Yiding Wang: resources and funding acquisition. Frank K. Tittel: writing – review and editing and supervision.

Conflicts of interest

There are no conflicts to declare.

Acknowledgements

This work was supported in part by the National Key R&D Program of China (no. 2021YFB3201903), the National Natural Science Foundation of China (nos 61960206004 and 62175087), the Science and Technology Development Program of Jilin Province, China (no. 20200401059GX), the Science and Technology Research Project of Department of Education, Jilin Province, China (no. JJKH20211088KJ), the Key R&D Program of Changchun (no. 21ZGN24), and the Program for JLU Science and Technology Innovative Research Team (JLUSTIRT, 2021TD-39).

References

1. W. H. Zhou, S. D. Xv, C. J. Liu and J. F. Zhang, Appl. Spectrosc. Rev., 2016, 51, 298–312.
2. K. G. Fue, W. M. Porter, E. M. Barnes and G. C. Rains, AgriEngineering, 2020, 2, 150–174.
3. Y. L. Chen, H. Z. Yang and W. Z. Liu, The First International Conference on Information Technology and Quantitative Management, ITQM, Suzhou, China, 2013.
4. J. S. Tian, X. Y. Zhang, Y. L. Yang, C. X. Yang, S. Z. Xu, W. Q. Zuo, W. F. Zhang, H. Y. Dong, X. L. Jiu, Y. C. Yu and Z. Zhao, Ind. Crops Prod., 2017, 109, 803–811.
5. Y. J. Shen, S. Li, Y. N. Chen, Y. Q. Qi and S. W. Zhang, Agric. Water Manage., 2013, 128, 55–64.
6. N. Li, H. X. Lin, T. X. Wang, Y. Li, Y. Liu, X. G. Chen and X. T. Hu, Field Crops Res., 2020, 247, 107590.
7. T. Hu, Y. N. Hu, J. Q. Dong, S. J. Qiu and J. Peng, Remote Sens., 2021, 13, 4819.
8. E. Mygdakos and T. A. Gemtos, Bioprocess Eng., 2002, 82, 381–391.
9. M. V. Braunack and D. B. Johnston, Soil Tillage Res., 2014, 140, 29–39.
10. Y. Q. Gu, H. W. Zhang, X. Q. Fu, L. Wang, Z. Y. Shen, J. Wang, Z. Y. Song and L. C. Zhang, Materials, 2021, 14, 2487.
11. Y. Q. Zhang, Y. Tian and Y. G. Meng, Trans. ASME, J. Tribol., 2021, 143, 021703.
12. G. S. Qi, W. Lu, X. Y. Qi, X. X. Zhong, W. M. Cheng and F. M. Liu, Fire Saf. J., 2018, 102, 77–82.
13. J. H. Du, R. C. Mang, X. Y. Huang, X. Gong and X. H. Zhang, Spectrosc. Spectrum Anal., 2007, 27, 899–903.
14. A. Solorzano, J. Eichmann, L. Fernandez, B. Ziems, J. M. Jimenez-Soto, S. Marco and J. Fonollosa, Sens. Actuators, B, 2021, 352, 130961.
15. D. Gutmacher, U. Hoefer and J. Woellenstein, Sens. Actuators, B, 2013, 175, 40–45.
16. Q. X. He, C. T. Zheng, H. F. Liu, Y. D. Wang and F. K. Tittel, Quantum Sensing and Nano Electronics and Photonics XIV, San Francisco, USA, 2017.
17. W. L. Ye, C. T. Zheng, X. Y. Cheng, F. P. Wu, M. M. Chen and Y. D. Wang, Infrared Phys. Technol., 2015, 71, 339–346.
18. Y. Chi, Z. X. Liu and Y. M. Zhang, J. Intell. Robot. Syst., 2019, 93, 337–349.
19. J. M. Dang, H. Y. Yu, C. T. Zheng, Y. D. Wang and Y. J. Sun, Anal. Methods, 2018, 10, 3325–3331.
20. J. W. Wang and H. L. Wang, Infrared Phys. Technol., 2014, 65, 1–4.
21. N. Pakmanesh, S. M. Cristescu, A. Ghorbanzadeh, F. J. M. Harren and J. Mandon, Appl. Phys. B: Lasers Opt., 2016, 122, 10.
22. S. Viciani, A. Montori, A. Chiarugi and F. D'Amato, Sensors, 2018, 18, 2380.
23. J. S. Li, H. Deng, J. Sun, B. L. Yu and H. Fischer, Sens. Actuators, B, 2016, 231, 723–732.
24. N. W. Liu, L. G. Xu, S. Zhou, L. Zhang and J. S. Li, ACS Sens., 2020, 5, 3607–3616.
25. J. S. Li, U. Parchatka and H. Fischer, Sens. Actuators, B, 2013, 182, 659–667.
26. R. M. Spearrin, C. S. Goldenstein, I. A. Schultz, J. B. Jeffries and R. K. Hanson, Appl. Phys. B: Lasers Opt., 2014, 117, 689–698.
27. C. Liu and L. J. Xu, Appl. Spectrosc. Rev., 2019, 54, 1–44.
28. N. W. Liu, L. G. Xu, S. Zhou, L. Zhang and J. S. Li, Analyst, 2021, 146, 3841.
29. H. Hussain, J. H. Kim and S. H. Yi, Sensors, 2018, 18, 2911.
30. M. M. Akram, A. Nikfarjam, H. Hajghassem, M. Ramezannezhad and M. Iraj, Sens. Rev., 2020, 40, 637–646.
31. T. A. Vincent and J. W. Gardner, Sens. Actuators, B, 2016, 236, 954–964.
32. S. J. Chen, D. C. Hovde, K. A. Peterson and A. W. Marshall, Fire Saf. J., 2007, 42, 507–515.
33. Z. J. Xie and Q. L. Tan, Int. J. Infrared Millimeter Waves, 2006, 27, 1639–1649.
34. J. N. Wang, L. J. Zheng, X. T. Niu, C. T. Zheng, Y. D. Wang and F. K. Tittel, Appl. Opt., 2016, 55, 7029–7036.
35. M. Dong, C. T. Zheng, S. Z. Miao, F. Song and Y. D. Wang, Infrared Phys. Technol., 2017, 85, 450–456.
36. Z. W. Liu, C. T. Zheng, C. Chen, H. T. Xie, Q. Ren, W. L. Ye, Y. D. Wang and F. K. Tittel, Anal. Methods, 2018, 10, 4838–4844.
37. F. Song, C. T. Zheng, D. Yu, Y. W. Zhou, W. H. Yan, W. L. Ye, Y. Zhang, Y. D. Wang and F. K. Tittel, Appl. Phys. B: Lasers Opt., 2018, 124, 51.

---